Enhancing Sentiment Analysis Accuracy and Evaluating Task Affinity Using Large Language Models

Sentiment analysis is a crucial natural language processing task that involves extracting emotions and opinions from text data, with applications spanning marketing, social media analysis, and enhancing user experience. Traditional approaches to sentiment analysis, such as dictionary-based and machine learning-based methods, have been predominant; however, these methods have limitations in handling complex emotional expressions and contextual dependencies. Recently, the advent of Large Language Models (LLMs) has raised expectations for improving sentiment analysis accuracy, and research exploring this potential is ongoing. In this study, we employed the Llama 2-7B model, an LLM, to thoroughly evaluate the effects of Supervised Fine-Tuning (SFT) and Few-shot Learning on various sentiment analysis subtasks. Specifically, we focused on tasks such as Sentiment Classification (SC), Aspect-Based Sentiment Analysis (ABSA), and Multifaceted Analysis of Subjective Text (MAST) to assess the affinity of LLMs for these tasks. The experimental results demonstrated that Few-shot Learning significantly improved accuracy across many tasks, with the 2-shot setting proving particularly effective. On the other hand, SFT led to a decrease in accuracy for certain tasks, suggesting the need for careful design in LLM training methods. Additionally, it was revealed that prompt design considering contextual information contributes to performance improvements in tasks such as irony and offensive language detection. These findings provide valuable insights into the design and training methods for specializing LLMs in sentiment analysis, offering important guidance for future research and practical applications.